U.S. patent number 8,595,313 [Application Number 11/289,140] was granted by the patent office on 2013-11-26 for systems and method for simple scale-out storage clusters.
This patent grant is currently assigned to NetApp. Inc.. The grantee listed for this patent is William P. Delaney, Mohamad H. El-Batal, Bret S. Weber. Invention is credited to William P. Delaney, Mohamad H. El-Batal, Bret S. Weber.
United States Patent |
8,595,313 |
Weber , et al. |
November 26, 2013 |
Systems and method for simple scale-out storage clusters
Abstract
Systems and associated methods for flexible scalability of
storage systems. In one aspect, a storage controller may include an
interface to a fabric adapted to permit each storage controller
coupled to the fabric to directly access memory mapped components
of all other storage controllers coupled to the fabric. The CPU and
other master device circuits within a storage controller may
directly address memory an I/O devices directly coupled thereto
within the same storage controller and may use RDMA features to
directly address memory an I/O devices of other storage controllers
through the fabric interface.
Inventors: |
Weber; Bret S. (Wichita,
KS), El-Batal; Mohamad H. (Westminster, CO), Delaney;
William P. (Wichita, KS) |
Applicant: |
Name |
City |
State |
Country |
Type |
Weber; Bret S.
El-Batal; Mohamad H.
Delaney; William P. |
Wichita
Westminster
Wichita |
KS
CO
KS |
US
US
US |
|
|
Assignee: |
NetApp. Inc. (Sunnyvale,
CA)
|
Family
ID: |
38088791 |
Appl.
No.: |
11/289,140 |
Filed: |
November 29, 2005 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20070124407 A1 |
May 31, 2007 |
|
Current U.S.
Class: |
709/212;
709/214 |
Current CPC
Class: |
G06F
3/0658 (20130101); H04L 67/1097 (20130101); G06F
12/0866 (20130101); G06F 15/17331 (20130101); G06F
3/067 (20130101); G06F 3/0607 (20130101) |
Current International
Class: |
G06F
15/167 (20060101) |
Field of
Search: |
;709/212 |
References Cited
[Referenced By]
U.S. Patent Documents
Primary Examiner: Towfighi; Afshawn
Attorney, Agent or Firm: Haynes and Boone, LLP
Claims
What is claimed is:
1. An apparatus within a storage controller for communication among
a plurality of storage controllers in a clustered storage system,
the apparatus comprising: a local memory mapped interface for
access to local cache memory of the storage controller by other
components within the storage controller; a fabric communication
interface for inter-controller communication between the storage
controllers and other storage controllers of the plurality of
storage controllers; and a remote memory mapped interface for
access to the local cache memory of the storage controller by the
other storage controllers, wherein the fabric communication
interface is adapted to permit remote DMA ("RDMA") access by the
storage controller to local cache memories of the other storage
controllers and is further adapted to permit RDMA access to the
local cache memory of the storage controller by the other storage
controllers.
2. The apparatus of claim 1 wherein the local memory mapped
interface includes: a map portion corresponding to each of the
other storage controllers to permit access by the storage
controller to the local cache memory of each of the other storage
controllers.
3. The apparatus of claim 2 wherein each map portion further
includes: an I/O map portion corresponding to one or more I/O
devices of the corresponding other storage controller to permit
RDMA access by the storage controller to the one or more I/O
devices of the corresponding other storage controller.
4. The apparatus of claim 1 wherein the remote memory mapped
interface further includes: an I/O map portion corresponding to one
or more I/O devices of the storage controller to permit RDMA access
by any of the other storage controllers to the one or more I/O
devices of the storage controller.
5. The apparatus of claim 1 wherein the storage controller performs
I/O requests by accessing information from the other storage
controllers using the fabric communication interface.
6. The apparatus of claim 1 wherein the storage controller
maintains cache coherency of local cache memories associated with
each of the plurality of storage controllers by exchanging cache
coherency information using the fabric communication interface.
7. A system comprising: a plurality of storage modules; a plurality
of storage controllers each adapted for coupling to one or more of
the plurality of storage modules and each adapted for coupling to
one or more host systems; and an inter-controller communication
medium coupling the plurality of storage controllers to one
another, wherein each storage controller includes: local cache
memory; a back-end interface for coupling the storage controller to
the one or more storage modules; a front-end interface for coupling
the storage controller to the one or more host systems; and an
inter-controller interface for coupling the storage controller to
one or more other storage controllers of the system through the
inter-controller communication medium, wherein the inter-controller
interface permits remote DMA ("RDMA") access between the storage
controller and the local cache memory of the one or more other
storage controllers and permits RDMA access to local cache memory
of the storage controller by the one or more other storage
controllers.
8. The system of claim 7 wherein the inter-controller communication
medium further comprises a parallel bus structure.
9. The system of claim 8 wherein the parallel bus structure is a
PCI Express bus structure.
10. The system of claim 7 wherein the inter-controller
communication medium further comprises a fabric communication
medium.
11. The system of claim 10 wherein the fabric communication medium
is an InfiniBand communication medium.
12. The system of claim 7 wherein each storage controller further
comprises: a first memory map that maps physical addresses within
the storage controller for access to the local cache memory, to the
back-end interface, and to the front-end interface; and a second
memory map that maps physical addresses within the storage
controller for RDMA access to local cache memory of the one or more
other storage controllers, to front-end interfaces of the one or
more other storage controllers, and to the back-end interface of
the one or more other storage controllers, wherein each storage
controller may access the local cache memory, front-end interface,
and back-end interface of any storage controller of the plurality
of storage controller.
13. The system of claim 7 wherein each storage controller is
adapted to perform I/O requests received from one or more host
systems by accessing information on the other storage controllers
using the inter-controller interface.
14. The system of claim 7 wherein each storage controller is
adapted to maintain cache coherency of local cache memories
associated with each of the plurality of storage controllers by
exchanging cache coherency information using the inter-controller
interface and the inter-controller communication medium.
15. A method operable in a storage controller of a storage system
having a plurality of storage controllers, the method comprising:
receiving an I/O request from a host system via a first
communication medium coupling the storage controller to one or more
host systems; responsive to receipt of the I/O request, accessing
information in one or more other storage controllers of the
plurality of storage controllers using remote DMA ("RDMA") via a
second communication medium coupling the plurality of storage
controllers to one another; and responsive to receipt of the I/O
request, accessing information on one or more storage modules via a
third communication medium coupling the storage controller to the
one or more storage modules.
16. The method of claim 15 wherein each of the plurality of storage
controllers includes a local cache memory and wherein the method
further comprises: maintaining coherency of the local cache memory
of the storage controller and the local cache memories of the other
storage controllers by exchanging cache coherency information
between the storage controller and the other storage controllers
via the second communication medium.
Description
BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention relates generally to storage clusters and more
specifically to systems and methods for simpler, low-cost scale-out
storage cluster configurations.
2. Related Patents
This patent application is related to co-pending, commonly owned
U.S. patent application Ser. No. 10/329,184 (filed Dec. 23, 2002;
the "'184 application" now published as Publication No.
2004/0123017), U.S. patent application Ser. No. 10/328,672 (filed
Dec. 23, 2002; now published as Publication No. 2004/0122987), and
U.S. patent application Ser. No. 10/671,158 (filed Sep. 25, 2003;
now published as Publication No. 2005/0071546), all of which are
hereby incorporated by reference. Additionally, U.S. Pat. No.
6,173,374 (issued Jan. 9, 2001) and U.S. Pat. No. 6,073,218 (issued
Jun. 6, 2000) provide useful background information and are hereby
incorporated by reference.
3. Discussion of Related Art
A mass storage system is used for storing user and system data in
data processing applications. A typical mass storage system
includes a plurality of computer disk drives configured for
cooperatively storing data as a single logically contiguous storage
space often referred to as a volume or logical unit. One or more
such volumes or logical units may be configured in a storage
system. The storage system, therefore, performs much like that of a
single computer disk drive when viewed by a host computer system.
For example, the host computer system can access data of the
storage system much like it would access data of a single internal
disk drive, in essence, oblivious to the substantially transparent
underlying control of the storage system.
Mass storage systems may employ Redundant Array of Independent
Disks ("RAID") management techniques, such as those described in "A
Case For Redundant Arrays Of Inexpensive Disks", David A. Patterson
et al., 1987 ("Patterson"). RAID levels exist in a variety of
standard geometries, many of which are defined by Patterson. For
example, the simplest array, a RAID level 1 system, comprises one
or more disks for storing data and an equal number of additional
"mirror" disks for storing copies of the information written to the
data disks. Other RAID management techniques, such as those used in
RAID level 2, 3, 4, 5, 6 and 10 systems, segment or stripe the data
into portions for storage across several data disks, with one or
more additional disks utilized to store error check or parity
information.
Regardless of storage management techniques, a mass storage system
may include one or more storage modules with each individual
storage module comprising a plurality of disk drives coupled to one
or more storage controllers. In one typical configuration, a
storage module may be coupled through its storage controller(s)
directly to a host system as a stand-alone storage module. Typical
storage controllers include significant cache memory capacity to
improve performance of the I/O operation. Write requests may be
completed when the supplied data is written to the higher speed
cache memory. At some later point, the data in cache memory may be
flushed or posted to the persistent storage of the storage modules.
Also, read requests may often be satisfied by accessing data
already resident in the higher speed cache memory of the storage
controller.
Such direct coupling of a storage system to a host system may
utilize any of numerous communication media and protocols. Parallel
SCSI buses are common for such direct coupling of a storage system
to a host. Fibre Channel and other high speed serial communication
media are also common in high performance environments where the
enterprise may require greater physical distance for coupling
between the storage systems and the host systems.
Even in a standalone configuration, it is common to enhance
reliability and performance by providing a redundant pair of
storage controllers. The redundant pair of controllers enhances
reliability in that an inactive storage controller may assume
control when the active controller is sensed to have failed in some
manner. The redundant pair of storage controller may also enhance
performance of the standalone storage system in that both storage
controller may be active--each acting as backup for the other while
both simultaneously processing different I/O requests or different
portions of an I/O request.
In such a configuration with redundant storage controllers, the
storage controllers typically exchange information to maintain
coherency of data between the cache memories resident in each
controller. Some prior storage systems use the communication path
between the controllers and the storage modules for the additional
cache coherency information exchanges. However, such shared use of
this communication path for interaction between the controller and
the storage modules and for cache coherency exchanges between the
storage controllers can negatively impact storage system
performance. Some prior techniques have therefore provided a
dedicated bus or channel coupled between the redundant pair of
storage controllers and intended primarily for such cache coherency
exchanges. Such a bus or dedicated communication channel is
typically adapted well for simple, fast, point-to-point exchanges
between the paired redundant storage controllers.
In another standard configuration, the storage module may be part
of a larger storage network or cluster. In a storage
network/cluster architecture, a plurality of storage modules and
corresponding storage controller are typically coupled through a
switched network communication medium (i.e., a fabric) to one or
more host systems. This form of a multiple storage module system is
often referred to as a Storage Area Network ("SAN") architecture
and the switching fabric is, therefore, often referred to as an SAN
switching fabric. In such a clustered configuration it is common
that all of the storage controllers exchange coherency information
as well as other information for load balancing of I/O request
processing and other control information. Such control information
may be exchanged over the same network fabric that couples the
storage controllers to the host systems (often referred to as the
"front end" connection or fabric) or over another fabric that
couples the storage controllers to the storage modules (often
referred to as the "back-end" connection or fabric). Though such a
fabric connection allows scalability of the storage controllers,
use of the existing front end or back-end fabric may negatively
impact overall storage system performance.
The differences between a stand-alone storage system and a storage
network architecture are marked. In a stand-alone storage module
system, a host computer system will directly send Input/Output
("I/O") requests to the storage controller(s) of the storage
module. The storage controller receiving the request, in general,
completely processes the received I/O requests to access data
stored within the disk drives of the storage module. The storage
controller then identifies and accesses physical storage spaces by
identifying and accessing particular logical unit numbers ("LUNs"
often also referred as "volumes" or "logical volumes") within one
or more of the disk drives of the storage module. Via the storage
controller, the host computer system can then read data from the
storage spaces or write data to the physical storage spaces.
By way of contrast, in a multiple storage module configuration
(i.e., networked storage or storage cluster), the various LUNs or
even a single LUN can be spread across one or more storage modules
of the storage system. In such a multiple module storage system the
switching fabric may be used to effectuate communication between
the storage controllers of one or more storage modules (e.g., via
the back-end fabric) as well as between the storage controllers and
the host systems (e.g., via the front end fabric). A host computer
may communicate an I/O request to the storage system and,
unbeknownst to the host system, the I/O request may be directed
through the switching fabric to any storage controller of any of
the storage modules. The storage controllers of multiple storage
modules may require communications for exchange of cache coherency
information and to coordinate and share information regarding LUNs
that are distributed over multiple storage modules. Information
returned by the storage controllers is routed back through the
switched fabric to the requesting host system.
For any of several reasons, an enterprise may wish to change from a
direct coupled storage module to a storage network/cluster
architecture for coupling storage modules to host systems. For
example, a network/cluster architecture may allow for increased
available communication bandwidth where multiple host communication
links may be available between the networked complex of storage
modules and one or more host systems. Another potential benefit of
a network/cluster storage architecture derives from the increased
storage performance realized by the cooperative processing of
multiple storage controllers that are interconnected to share the
workload of requested I/O operations. Another possible reason for
an enterprise to convert to a storage network/cluster architecture
is to increase storage capacity beyond the capacity of a single,
stand-alone storage module. The above-mentioned benefits and
reasons may hereinafter be collectively referred to as storage
performance features.
Any particular storage module has a finite storage capacity
because, for example, a storage module has a finite physical area
in which the disk drives may be placed. In addition, performance of
the storage module may be limited to a number of possible
controllers that may be configured within a stand-alone storage
module for processing of host system I/O requests. Alternatively, a
storage module may have a limit on the number of direct host
communication links and hence a limit on the available bandwidth
for communicating between the storage subsystem and host systems.
Accordingly, when an organization requires improved performance
features from its storage system, the organization may implement a
new storage system designed with multiple storage modules in a
storage network architecture to provide additional storage capacity
and/or performance to overcome the limitations of a single
stand-alone storage module.
Since a stand-alone storage module has a controller configured for
direct access by a host computer system but typically not for
cooperation and coordination with other controllers of other
storage modules, implementation of a new multiple storage module
networked storage system may include replacement of the storage
controller(s) of the stand-alone storage module(s). Different
storage controllers may be required to provide the required
interconnection between storage controllers of the multiple storage
modules to permit desired cooperation and coordination between the
multiple storage modules. Such a reconfiguration of the stand-alone
storage module is necessary because the storage module may
coordinate with other storage modules through an SAN switching
fabric not previously required in a stand-alone storage module.
Upgrades to an existing stand-alone storage system to enable
networked communications among multiple storage modules remain an
expensive process. In addition to possible replacement of storage
controllers, retrofitting a present stand-alone storage module to
operate as one of a plurality of storage modules in a networked
storage system typically requires other components to implement
communication between the storage controllers. Costly, complex
N-way fabric switches add significant cost for the initial
conversion from a stand-alone configuration to a storage network
configuration.
Although storage performance feature requirements often grow in an
enterprise, the cost for conversion to a networked storage
architecture may be prohibitive to smaller enterprises. It is
therefore evident that a need exists to provide improved methods
and structure for improving storage performance feature scalability
to permit cost effective growth of storage as an organization
grows.
SUMMARY OF THE INVENTION
The present invention solves the above and other problems, thereby
advancing the state of the useful arts, by providing methods and
structures that enable flexible scaling of a storage subsystem from
a single storage controller, to a redundant pair of mirrored
controller and on to a full clustered storage environment with
N-way connectivity among any number of controllers in the
subsystem. This flexible scalability is provided without imposing
significant complexity by providing flexible configuration of any
pair of controllers as well as an N-way connection among all
controllers through a switched fabric inter-controller
connection.
A first feature hereof provides an apparatus within a storage
controller for communication among a plurality of storage
controllers in a clustered storage system, the apparatus
comprising: a local memory mapped interface for access to local
cache memory of the storage controller by other components within
the storage controller; a fabric communication interface for
inter-controller communication between the storage controllers and
other storage controllers of the plurality of storage controllers;
and a remote memory mapped interface for access to the local cache
memory of the storage controller by the other storage controllers,
wherein the fabric communication interface is adapted to permit
remote DMA ("RDMA") access by the storage controller to local cache
memories of the other storage controllers and is further adapted to
permit RDMA access to the local cache memory of the storage
controller by the other storage controllers.
Another aspect hereof further provides that the local memory map
includes: a map portion corresponding to each of the other storage
controllers to permit access by the storage controller to the local
cache memory of each of the other storage controllers.
Another aspect hereof further provides that each map portion
further includes: an I/O map portion corresponding to one or more
I/O devices of the corresponding other storage controller to permit
RDMA access by the storage controller to the one or more I/O
devices of the corresponding other storage controller.
Another aspect hereof further provides that the remote memory map
further includes: an I/O map portion corresponding to one or more
I/O devices of the storage controller to permit RDMA access by any
of the other storage controllers to the one or more I/O devices of
the storage controller.
Another aspect hereof further provides that the storage controller
performs I/O requests by accessing information from the other
storage controllers using the fabric communication interface.
Another aspect hereof further provides that the storage controller
maintains cache coherency of local cache memories associated with
each of the plurality of storage controllers by exchanging cache
coherency information using the fabric communication interface.
Another feature hereof provides a system comprising: a plurality of
storage modules; a plurality of storage controllers each adapted
for coupling to one or more of the plurality of storage modules and
each adapted for coupling to one or more host systems; and an
inter-controller communication medium coupling the plurality of
storage controllers to one another, wherein each storage controller
includes: local cache memory; a back-end interface for coupling the
storage controller to the one or more storage modules; a front-end
interface for coupling the storage controller to the one or more
host systems; and an inter-controller interface for coupling the
storage controller to one or more other storage controllers of the
system through the inter-controller communication medium, wherein
the inter-controller interface permits remote DMA ("RDMA") access
between the storage controller and the local cache memory of the
one or more other storage controllers and permits RDMA access to
local cache memory of the storage controller by the one or more
other storage controllers.
Another aspect hereof further provides that the inter-controller
communication medium further comprises a parallel bus
structure.
Another aspect hereof further provides that the parallel bus
structure is a PCI Express bus structure.
Another aspect hereof further provides that the inter-controller
communication medium further comprises a fabric communication
medium.
Another aspect hereof further provides that the fabric
communication medium is an InfiniBand communication medium.
Another aspect hereof further provides that each storage controller
further comprises: a first memory map that maps physical addresses
within the storage controller for access to the local cache memory,
to the back-end interface, and to the front-end interface; and a
second memory map that maps physical addresses within the storage
controller for RDMA access to local cache memory of the one or more
other storage controllers, to front-end interfaces of the one or
more other storage controllers, and to the back-interface of the
one or more other storage controllers, wherein each storage
controller may access the local cache memory, front-end interface,
and back-end interface of any storage controller of the plurality
of storage controller.
Another aspect hereof further provides that each storage controller
is adapted to perform I/O requests received from one or more host
systems by accessing information on the other storage controller
using the inter-controller interface.
Another aspect hereof further provides that each storage controller
is adapted to maintain cache coherency of local cache memories
associated with each of the plurality of storage controllers by
exchanging cache coherency information using the inter-controller
interface and the inter-controller communication medium.
BRIEF DESCRIPTION OF THE DRAWINGS
FIGS. 1a through 1c are block diagrams of an exemplary system
including a storage controller in accordance with features and
aspects hereof to allow flexible scaling of controller through a
dedicated inter-controller interface.
FIG. 2 is a block diagram of another exemplary system including a
storage controller in accordance with features and aspects hereof
to allow flexible scaling of controller through a dedicated
inter-controller interface.
FIG. 3 is a block diagram of an exemplary local access memory map
and an exemplary remote access memory map as noted in FIG. 2 in
accordance with features and aspects hereof.
FIG. 4 is a flowchart describing an exemplary method for processing
an I/O request in accordance with features and aspects hereof.
FIG. 5 is a block diagram of an exemplary storage system scaled to
a large scale subsystem using a fabric connection among the
plurality of storage controllers in accordance with features and
aspects hereof.
DETAILED DESCRIPTION OF THE DRAWINGS
FIG. 1a is a block diagram of a system 100 including storage
controllers 102 coupled via communication path 150 to storage
modules 104. As generally known in the art, storage modules 104 may
comprise individual disk drives or enclosures of multiple disk
drives incorporating one or more power supplies and cooling
systems. Such storage modules 104 may be coupled to storage
controllers 102 by path 150 using any of several well-known,
commercially available communication media and associated protocols
including, for example, parallel SCSI, serial attached SCSI
("SAS"), Fibre Channel, and various other parallel bus and high
speed serial communication architectures. Further, as generally
known in the art, path 150 may represent multiple redundant
communication paths to further improve reliability of
communications between a storage controller 102 and associated
storage modules 104.
Storage controller 102 may include an I/O request processor 106. In
general, I/O request processor 106 may include a general or special
purpose CPU 101 with associated program memory for controlling
overall operation of the storage controller. Back-end interface 110
within I/O request processor 106 provides interface features to
couple the storage controller 102 via communication path 150 to
attached storage modules 104. Front-end interface 108 within I/O
request processor 106 provides front-end interface capabilities to
couple storage controller 102 to one or more attached host systems
120. I/O request processor 106 generally includes high speed DMA
(direct memory access) capabilities to permit rapid transfer of
information between the various components of I/O processor 106 and
between the elements of I/O processor 106 and an associated local
cache memory 112. Path 152 represents typical local DMA
capabilities wherein back-end interface 110 may utilize DMA
features to transfer data between local cache memory 112 and
storage modules 104. In like manner, front-end interface 108 may
use local DMA features to transfer the information between local
cache memory 112 and attached host systems 120.
Storage controller 102 may also include inter-controller interface
element 114 to provide a dedicated communication path between
multiple storage controllers 102. The front-end communication path,
the back-end communication path and the inter-controller
communication path are preferably separate and distinct
communication paths. As noted above, such a dedicated communication
path may be applied for cache coherency information exchanges, I/O
request processing information exchanges, and any other
inter-controller exchanges useful for high performance I/O request
processing distributed among the plurality of storage controllers
102. Storage controller 102 is therefore adapted to permit local
processing of received I/O requests through operation of I/O
request processor 106 utilizing local DMA features via path 152 to
exchange information between attached host systems 120 and local
cache memory 112 and/or between storage modules 104 and local cache
memory 112. Storage controller 102 is also adapted to utilize
inter-controller interface 114 to exchange information with other
storage controllers in a clustered storage environment. As noted
above and as discussed further herein below, inter-controller
interface 114 provides a remote DMA ("RDMA") interface between
elements of storage controller 102 and elements of other similar
controllers within a storage cluster environment.
By contrast with current storage controller architectures, storage
controller 102 of FIG. 1a is easily scaled from a stand-alone
storage system having a single storage controller to a complex
storage cluster environment having multiple storage controllers
coupled together through their respective inter-controller
interface elements 114. The scaling of the system may be flexibly
scaled from a single, stand-alone storage controller to a redundant
pair of controllers or to multiple pairs of controllers where a
dedicated mirroring channel may still be used for cache mirroring.
The dedicated channel used for mirroring may also be used for other
inter-controller communication and/or another fabric interface may
be used for such other inter-controller k communication. Further,
the controller may scale to a larger topology by using the fabric
interface for cache mirroring and for other inter-controller
communications.
Thus a storage system in accordance with features and aspects
hereof may scale up in pairs or may scale up incrementally such
that mirroring and redundancy pairing may be configured dynamically
among the plurality of controllers in the scaled up store system.
The flexible scaling features hereof using a fabric connection
among all controllers also enhance performance of the scaled up
system by allowing dynamic adjustment of I/O processing loads since
any controller may mirror operations of another. In such an
environment, a feature hereof provides that the inter-controller
coupling utilizes RDMA capabilities for each controller to access
components of another controller coupled through the
inter-controller communication paths.
FIG. 1b shows the same storage controller architecture of FIG. 1
but highlights communication paths 154 and 156 that provide remote
DMA capabilities from I/O request processor 106 through the
inter-controller interface 114 to other storage controllers via
path 156. Utilizing these RDMA communication paths and
inter-controller interface 114, I/O request processor 106 may
transfer information from its local cache memory to elements within
another storage controller through inter-controller interface 114
and communication path 156. In like manner, I/O request processor
106 may utilize RDMA features to transfer information directly from
front-end interface 108 or back-end interface 110 through interface
controller 114 and communication path 156 to elements within other
storage controllers. In effect, the I/O devices (front-end
interface 108 and back-end interface 110) and local cache memory
112 of each of multiple storage controllers 102 may be directly
accessed by any other storage controller in a cluster storage
environment utilizing inter-controller interface 114 of each of the
storage controllers 102. Thus, system 100 of FIGS. 1a and 1b may be
easily and flexibly scaled from a single controller architecture
through any number of storage controllers. RDMA features and
aspects hereof permit each controller to access I/O devices and
local cache memory of any other storage controller in the cluster
configuration.
To enable use of RDMA features, the plurality of controllers
maintain memory maps such that each controller may access cache
memory and/or I/O interface devices of any other controller through
the inter-controller fabric communication links. In one aspect, a
master memory map may be created (i.e., at system initialization or
as a function of initial configuration of he system). Such a common
shared memory map is then used by each of the plurality of storage
controllers and may be maintained as updates are required by
exchanging update messages among the controllers. In another
aspect, each storage controller may construct its own memory map to
map to memory and I/O devices of each other controller that it may
access.
Those of ordinary skill in the art will recognize a wide variety of
communication media and protocols useful for the inter-controller
interface 114 to couple such a plurality of storage controllers
102. For example, a simple PCI Express bus providing RDMA
capabilities may be useful for coupling a small number of storage
controllers in a small cluster environment. Or, for example, high
speed serial communication protocols and media such as InfiniBand
may be useful for inexpensively coupling a larger number of storage
controllers in a larger, more complex storage cluster environment.
Still further, those of ordinary skill in the art will recognize
that a storage controller 102 may incorporate multiple such
communication features within inter-controller interface 114. For
example, inter-controller interface 114 may include both a PCI
Express bus interface and an InfiniBand interface to permit
flexible reconfiguration of the storage controller for a very small
cluster storage environment or a larger, complex storage cluster
environment.
As noted above, a storage controller in accordance with features
and aspects hereof may utilize a single communication link for both
mirrored cache information exchange as well as for other
inter-controller communication. Either a dedicated, point to point
communication channel such as a cache mirror channel may be used
for both purposes or a switched fabric channel may be used for both
communication purposes. To further enhance performance, a plurality
of storage controllers in accordance with features and aspects
hereof may use both a mirrored cache communication channel for
mirrored cache information exchanges and a switched fabric
communication link for exchange of other inter-controller
information.
FIG. 1c shows yet another usage of the communication paths utilized
in storage controller 102 in accordance with features and aspects
hereof. Another storage controller may access the I/O devices and
local cache memory of the depicted storage controller 102 using
RDMA capabilities. An external storage controller (not shown) may
utilize RDMA capabilities via path 156 through inter-controller
interface 114 to access front-end interface 108, back-end interface
110, and local cache memory 112 of the depicted storage controller
102. As noted above, such RDMA capabilities allow the external
storage controller (not shown) to easily manage cache coherency
between its local cache memory and the local cache memory 112 of
the depicted storage controller 102. RDMA capabilities allow an
external storage controller to directly access front-end interface
108 and back-end interface 110 (thereby permitting access to
storage modules 104 coupled through back-end interface 110). Thus,
the external storage controller may process I/O requests utilizing
the I/O devices, local cache memory, and storage modules associated
with the depicted storage controller 102.
Those of ordinary skill in the art will readily recognize that the
storage controller 102 of FIGS. 1a through 1c are intended merely
as exemplary of one possible implementation of features and aspects
hereof. Numerous additional components and features for a fully
functional storage controller will be readily apparent to those of
ordinary skill and the art. Further, those of ordinary skill in the
art will readily recognize that a variety of levels of integration
or further separation of components within a storage controller may
be selected as a matter of design choice. For example, the general
or special purpose processing features of the depicted CPU 101 of
storage controller 102 may be integrated within a single integrated
circuit with the front-end and back-end interface elements 108 and
110 as well as associated program memory. Such matters of design
choice are well known to those of ordinary skill in the art.
FIG. 2 shows a system 200 depicting other features and aspects
hereof wherein multiple storage controllers utilize RDMA
capabilities to cooperatively process I/O requests. One or more
host systems 202 are coupled to storage controller 204 to access to
storage modules 230. As above, storage modules 230 may be
individual disk drives or modules/enclosures housing multiple disk
drives or other storage devices coupled along with associated
common power supplies and cooling components--often in a redundant
configuration for enhanced reliability and performance. Host
systems 202 and storage modules 230 are coupled to storage
controller 204 through appropriate I/O interfaces and associated
communication paths and media. Exemplary of such interfaces and
communication paths and media are Fibre Channel, serial attached
SCSI ("SAS"), parallel SCSI, and other well-known parallel bus
structures and high speed serial communication structures and
protocols.
Storage controller 204 may include I/O request processing element
212 for receiving an I/O request from an attached host system 202
and processing the received requests by accessing identified
storage locations in storage modules 230. I/O request processing
element 212 may process received I/O requests by accessing
information in local cache memory and interfaces 214 or by
accessing local cache memory and interfaces 234 within another
storage controller 224. To access local cache memory and interfaces
214 within the same storage controller 204, I/O request processing
element 212 utilizes local access memory map 210 to address local
cache memory and interfaces 214. The memory map may provide well
known virtual memory addressing features to map a logical address
to the local physical address of the desired cache memory or I/O
device. Thus, an I/O request may be locally processed within
storage controller 204 by directly accessing identified locations
of storage modules 230 or of local cache memory.
Alternatively, storage controller 204 may process an I/O request by
accessing local cache memory and interfaces 234 associated with
another storage controller 224 in a clustered storage environment.
Other storage controller 224 may be a redundant controller paired
with storage controller 204 or may be any other controller in a
clustered storage environment providing a plurality of storage
controllers. Storage controller 204 and other storage controllers
224 may be coupled through respective fabric communication
interfaces 206 and 226. Fabric communication interfaces 206 and 226
may provide, for example, PCI Express, InfiniBand, or other
commercially available communication media and protocols that
permit remote DMA ("RDMA") access therethrough. In order to access
local cache memory and interfaces 234 on another storage controller
224, I/O request processing element 212 of storage controller 204
utilizes a remote access memory map 208 for purposes of generating
RDMA operations through the fabric communication interface 206
destined for cache memory and interfaces 234 on another storage
controller 224. Thus, storage controller 204 may perform processing
for a received I/O request utilizing RDMA capabilities to access
storage modules 230 through another storage controller 224. System
200 therefore provides flexible scalability for a storage subsystem
including any number of storage controllers coupled in redundant
pairs or coupled in any manner such that any controller may access
cache memory an I/O interface devices of any other controller in
the system using RDMA access.
Those of ordinary skill in the art will recognize that any number
of storage controllers may be coupled in such a manner including,
pair-wise coupling of redundant pairs or fabric oriented coupling
of any number of storage controllers in the system. FIG. 2 is
therefore intended merely as exemplary of one possible system
configuration employing features and aspects hereof to permit
flexible reconfiguration of any number of storage controllers in
the system 200.
FIG. 5 shows an exemplary storage system 500 scaled up to four
controllers (502. 504, 506, and 508) using a switched fabric 510
for inter-controller communication (including cache mirroring
exchanges as well). A second switched fabric 514 may couple the
controllers to one or more host systems 516. Thus any host system
516 may communicate to the storage system 500 through any of the
storage controllers 502 . . . 508 to transmit I/O requests and to
receive results including requested data. The controllers 502 . . .
508 may exchange information, including RDMA transfers of
information, via switched fabric 510 through the heavier dashed
line communication paths of FIG. 5. Thus any controller 502 . . .
508 may serve as a redundant backup for any other controller.
Further, any controller may access cache memory and/or I/O devices
of any other controller using RDMA transfers through the switched
fabric 510. In addition, or in the alternative, a heavier dotted
line in FIG. 5 represents a cache mirroring communication channel.
Such a mirrored cache communication channel may be used exclusively
for exchange of mirroring information among the redundant
controllers while the inter-controller communication may be
directed through the switched fabric 510. Such a configuration
enhances performance by segregating all controller interaction
while reducing cost and complexity of each controller. Thus a
single, simple controller design may be used to scale from a single
storage controller in a storage system up to any number of
controllers exchanging control information and exchanging mirrored
cache information.
The switched fabric 410 may be, as noted herein, an InfiniBand
fabric as well as numerous other high speed serial communication
protocols and media. Further, current proposals to advance the
design of the PCI Express bus architecture may extend this
communication medium to allow switched fabric architecture
communications. In particular, the "advanced switching
interconnect" special interest group has begun work as an industry
standards organization to define such a switched fabric-like
extension to the PCI Express structures and protocols. Those of
ordinary skill in the art will be aware of such developing
proposals and information regarding such developments is readily
available at www.asi-sig.org.
Still further, the enhanced architectures diagramed in FIGS. 1a . .
. 1c, 2, and 5 may be operated in accordance with a variety of
modes of operation to perform received I/O requests. The
architecture permits I/O requests to be performed using "I/O
shipping" techniques well known to those of ordinary wherein a
first storage controller receives an I/O request from an attached
host and transfers ("ships") the request to another storage
controller for further processing. In preferred modes of operation
and in accordance with features and aspects hereof, a first control
may receive an I/O operation. The first controller may process the
I/O request locally by accessing cache memory and I/O interfaces to
storage devices local to that first controller. In addition, the
first controller may utilize RDMA capabilities of the controller in
accordance with features and aspects hereof to access cache memory
and/or I/O interfaces of another storage controller in a cluster
coupled through a fabric connection to the first controller. In yet
another alternative preferred mode of operation, a first controller
may receive an I/O request but may "ship" the request to another
controller (e.g., to balance processing loads among the various
storage controllers in a cluster environment). The second
controller may then process the I/O request by using RDMA
techniques through a fabric connection to the other controllers. In
particular, the second controller may use RDMA features and aspects
hereof to access the cache memory and/or I/O interface devices of
he first controller--i.e., the controller that shipped the request
to this second controller.
In effect, in a clustered storage subsystem where all controllers
are coupled through a switched fabric interface, the processor of
any of the storage controller may access remote cache memory and
I/O interface devices of other controller in essentially he same
way that it may access its own local cache memory and I/O interface
devices. Thus features and aspects hereof provide improved
flexibility in the configuration of clustered storage systems and
improvements in performance of such a clustered storage system.
Further, these benefits are achieved in a scalable controller
design that may easily and inexpensively be scaled from a single,
stand-alone storage controller to a large clustered storage system
coupling any number of storage controllers.
In typical operation of clustered, scalable controllers in
accordance with features and aspects hereof, DMA scatter/gather
lists ("S/G lists") are passed from a first controller to a second
controller to provide local memory addresses within the first
controller's cache memory for a second controller to access using
RDMA features. In a read operation, the S/G list may represent
locations in the first controller's cache memory where data read
from storage device may be stored for eventual return to a
requesting host system. Or, the S/G list entries in a read
operation may represent locations in the local cache memory where
requested data is already stored such that the second controller
may return the data to the requesting host system. In like manner,
when performing a write operation, the S/G list may indicate local
cache memory locations in a first controller that a second
controller may access to record the data on the storage
devices.
FIG. 3 provides additional details of exemplary memory maps
utilized by system 200 of FIG. 2 for accessing either local cache
memory an I/O interfaces within the same storage controller or for
accessing cache memory an I/O interface controllers of any other
storage controller in a clustered storage system. Local memory map
210 is used by the I/O request processor of a storage controller to
gain access to the locally memory mapped I/O devices and local
cache memory of the storage controller that received the I/O
request. A first portion 320 of the local access memory map 210
maps logical memory addresses to I/O devices of this controller.
This first portion 320 is used by the I/O request processor to
access, for example, front-end interface elements and back-end
interface elements for exchanging information with the attached
host systems and attached storage modules, respectively. A second
portion 322 of local access memory map 210 maps logical memory
addresses to the local cache memory of this controller. The I/O
request processor of this controller uses these mapped memory
addresses for accessing information in the local cache memory of
this controller.
By contrast, remote access memory mapped 208 is used by the I/O
request processor of a controller to access I/O devices and/or
local cache memory of other controllers in the clustered storage
system. The remote access memory map 208 may include a sub-map or
portion associated with each other controller coupled to the
controller containing the remote access map 208. A first portion
300 identifies memory mapping information for a first other
controller "A" while the second and third controller "B" and "C",
respectively, have corresponding other portions 306 and 312,
respectively. Within each portion associated with a corresponding
other controller, a first portion 302 of the remote map identifies
logical memory addresses to be used by the current controller to
access I/O devices in the other controller "A". A second portion
304 identifies memory addresses to be used by the current storage
controller to access local cache memory of the other controller
"A". In like manner, portion 306 for controller "B" includes a
first portion 308 for mapping I/O devices of controller "B" and a
second portion 310 for mapping local cache memory of other
controller "B". Similarly, portion 312 includes a first portion 314
for mapping addresses used by the current controller to access I/O
devices of other controller "C" and a second portion 316 to access
local cache memory of other controller "C".
The current controller therefore uses the local access memory map
210 when processing an I/O request local within that controller.
I/O devices and local cache memory are accessed by corresponding
logical addresses identified in the respective portions 320 and 322
of local access memory map 210. The current controller may also
process an I/O request by accessing local cache memory and/or I/O
devices corresponding with another controller coupled to the
current controller. When so processing an I/O request, the current
controller uses the appropriate portion of the remote access memory
mapped 208 corresponding to the particular other controller and the
particular devices and/or local cache memory to be accessed.
Those of ordinary skill in the art will readily recognize numerous
detailed structures and alternative structures for representing the
memory mapped features shown in FIG. 3. In addition, memory
management features associated with various general and special
purpose processors may provide for specific register structures and
memory structures (e.g., within associated memory controller
circuits or chips) for retaining the mapping information associated
with the local I/O devices and local cache memory and the I/O
devices an cache memory associated with other controllers coupled
to the current controller. As noted above, the current controller
preferably uses RDMA capabilities to simplify access to the remote
devices and cache memory associated with other storage controllers
coupled to the current controller.
FIG. 4 is a flowchart describing a method in accordance with
features and aspects hereof operable within a storage controller in
a storage cluster including multiple storage controllers coupled
through a switched fabric inter-controller communication medium. As
noted above, when scaled up to such a cluster configuration, a
switched fabric communication path may be utilized to distribute
processing of I/O requests among the plurality of storage
controllers utilizing RDMA capabilities. Element 400 of FIG. 4 is
first operable to receive an I/O request from an attached host
system through a front-end interface of this storage controller.
Element 402 is then operable to determine whether the received I/O
request is preferably performed locally utilizing local I/O
interfaces (e.g., back-end interfaces to the storage devices) and
local cache memory of this storage controller (the controller in
receipt of the I/O request). Any number of factors well known to
those of ordinary skill in the art may be involved in such a
determination. Simple load balancing factors may be involved in a
decision to perform the received I/O request locally or to perform
the I/O request using capabilities of other remote storage
controllers in the storage cluster. Still further, the decision to
perform an I/O request locally may be made in accordance with a
determination that requested information to be access by the I/O
request is more readily available in the local cache memory of this
storage controller. Or the request may preferably be processed by
this controller if the eventual destination storage devices
associated with the I/O request are more easily accessible through
this storage controller. Those of ordinary skill in the art will
readily recognize numerous factors to be considered in such a
decision to process the I/O request totally locally. Where the
decision is made to perform the I/O request utilizing exclusively
local processing capabilities, element 404 represents normal I/O
processing within the storage controller to complete the I/O
request. Such normal processing generally entails utilization of
local cache memory within this storage controller and access to
storage devices through the back-end interface elements of this
storage controller.
Where another storage controller is better suited for any reason to
perform the received I/O request or to aid in completing the
request, element 406 is then operable to determine whether the
received request is for a write operation or for a read operation.
If the request is for a write operation, element 408 is next
operable to prepare an appropriate scatter/gather list for the
received data to be read in from the local cache memory of this
storage controller and written to appropriate storage devices. As
noted above, the information required to perform an I/O request on
another controller may preferably be provided to the other
controller in the form of a scatter/gather list indicating
addresses in the local cache memory of this first storage
controller where the data may be accessed. Element 410 is operable
to forward the write operation to another controller along with the
constructed scatter/gather list. Element 412 then represents
processing by the a selected other controller to perform the
indicated write operation using the scatter/gather list and using
RDMA capabilities to access the write data from the local cache
memory of this storage controller. As noted above, such RDMA
capabilities may be performed utilizing the inter-controller
interface through a switched fabric communication structure.
Lastly, element 414 is operable to return an appropriate completion
status to the requesting host system to thereby complete the
received I/O operation.
Where element 406 determines of that the received I/O request is
for a read operation, element 420 is next operable to determine
whether the requested data is resident in the cache memory one or
more other storage controllers in the storage cluster. Distributed
cache management techniques utilizing the inter-controller switched
fabric communication features and aspects hereof may readily
determine that requested data resides in the cache memory local to
another storage controller in the storage cluster. If element 420
determines that the requested data is so resident in the cache
memory of another controller, element 422 is operable to access the
requested information from one or more other controllers. RDMA
capabilities are used to access the requested information from the
cache memory of another controller using the inter-controller
switched fabric communication features and aspects hereof.
Those of ordinary skill in the art will recognize that the
information to be accessed may be received utilizing RDMA from this
controller to access the cache memory of one or more other
controllers in the storage cluster. Similarly, those of ordinary
skill in the art will recognize that the requested information may
also be first copied to the local cache memory of this storage
controller from the cache memory of another storage controller and
then returned to the requesting host system. In such an operation,
a scatter/gather list may be prepared and forwarded to the other
controllers in which the requested memory requested information is
resident in the cache memory. The other controller may then utilize
RDMA capabilities to transfer the requested data into the local
cache memory of this storage controller utilizing the
inter-controller switched fabric communication features and aspects
hereof. Numerous other equivalent transfers will be readily
apparent to those of ordinary skill in the art utilizing switched
fabric communications features and aspects hereof and RDMA
capabilities over such a switched fabric communication path.
Lastly, element 414 is operable as discussed above to complete the
I/O request by return of an appropriate completion status to the
requesting host system along with any requested read data.
If element 420 determines that the requested data is not resident
in the cache memory of any storage controller of the storage
cluster, element 424 is operable to "ship" the I/O request to
another controller that may more readily access the information
from associated storage devices. As generally known in the art, the
storage devices may be partitioned such that particular storage
controllers have access to particular storage devices or regions of
particular storage devices. Thus, the I/O request may be more
easily performed by a storage controller that has direct access to
the appropriate storage devices containing the requested
information. Element 424 therefore ships the I/O request to an
appropriate other controller and may await completion of that
request. Element 414 then returns an appropriate completion status
along with requested information to the requesting host system. In
the alternative, the controller to which the I/O request was
shipped may directly communicate with the requesting host system to
complete the I/O read operation. In such a context, this controller
may continue operation without awaiting completion of the shipped
I/O request.
The flowchart of FIG. 4 is intended merely as exemplary of one
method in accordance with features and aspects hereof to utilize
the cluster of storage controllers coupled through and
inter-controller switched fabric communication medium. Where such a
controller is scaled up to a large storage cluster, the method of
FIG. 4 allows for exchange of information among the clustered
storage controllers such that any of the storage controllers may
complete the I/O request utilizing the switched fabric
communication medium and RDMA capabilities thereon to optimally
complete the requested I/O operation. The same controller structure
when scaled down to a single controller architecture or a simple
dual redundant pair of controllers may perform all I/O operations
locally or through simple mirrored operation with its redundant
paired controller. Thus the structures and methods in accordance
with features and aspects hereof may easily scale from a single
storage controller context to a larger clustered storage controller
context where the cluster of storage controllers communicate
through the switched fabric communication medium.
Those of ordinary skill and the art will recognize a variety of
equivalent methods and detailed processing for completing
processing of a received I/O request within a storage controller
coupled to one or more other controllers in a storage cluster
environment. FIG. 4 is therefore merely intended as representative
of one possible embodiment of such a processing method.
While the invention has been illustrated and described in the
drawings and foregoing description, such illustration and
description is to be considered as exemplary and not restrictive in
character. One embodiment of the invention and minor variants
thereof have been shown and described. Protection is desired for
all changes and modifications that come within the spirit of the
invention. Those skilled in the art will appreciate variations of
the above-described embodiments that fall within the scope of the
invention. In particular, those of ordinary skill in the art will
readily recognize that features and aspects hereof may be
implemented equivalently in electronic circuits or as suitably
programmed instructions of a general or special purpose processor.
Such equivalency of circuit and programming designs is well known
to those skilled in the art as a matter of design choice. As a
result, the invention is not limited to the specific examples and
illustrations discussed above, but only by the following claims and
their equivalents.
* * * * *
References